init

.gitattributes

@@ -1,4 +1,8 @@
+data/**/* filter=lfs diff=lfs merge=lfs -text
 *.7z filter=lfs diff=lfs merge=lfs -text
+*.mdx filter=lfs diff=lfs merge=lfs -text
+*.pdf filter=lfs diff=lfs merge=lfs -text
+*.ttf filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.avro filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
@@ -58,3 +62,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+ocr/out/** filter=lfs diff=lfs merge=lfs -text
+ocr/out2/** filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,8 @@
+__pycache__/
+ocr/.venv/
+*.pdf
+!戚蓼生序本石头记.pdf
+!新刻金瓶梅词话.第001回至057回.总一百回.明.兰陵笑笑生撰.明万历四十五年刊本.台北故宫博物院藏.pdf
+!新刻金瓶梅词话.第058回至100回.总一百回.明.兰陵笑笑生撰.明万历四十五年刊本.台北故宫博物院藏.pdf
+
+

ocr/.vscode/launch.json

@@ -0,0 +1,12 @@
+{
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Python Debugger: Current File",
+            "type": "debugpy",
+            "request": "launch",
+            "program": "${file}",
+            "console": "integratedTerminal"
+        }
+    ]
+}

ocr/auto.py

@@ -0,0 +1,125 @@
+import subprocess
+import sys
+import os
+import shutil
+import time
+
+def main():
+    # 获取当前目录
+    current_dir = os.path.dirname(os.path.abspath(__file__))
+    getdata_script = os.path.join(current_dir, "getdata.py")
+    gendata_script = os.path.join(current_dir, "gendatav2.py")
+    
+    # 章节列表
+    chapters = [
+        "新刻金瓶梅詞話卷之十_第九十一回",
+        "新刻金瓶梅詞話卷之十_第九十二回",
+        "新刻金瓶梅詞話卷之十_第九十三回",
+        "新刻金瓶梅詞話卷之十_第九十四回",
+        "新刻金瓶梅詞話卷之十_第九十五回",
+        "新刻金瓶梅詞話卷之十_第九十六回",
+        "新刻金瓶梅詞話卷之十_第九十七回",
+        "新刻金瓶梅詞話卷之十_第九十八回",
+        "新刻金瓶梅詞話卷之十_第九十九回",
+        "新刻金瓶梅詞話卷之十_第一百回"
+    ]
+    
+    for i, chapter_name in enumerate(chapters):
+        print(f"\n=====================================")
+        print(f"开始处理章节: {chapter_name}")
+        print(f"=====================================")
+
+        if os.path.exists(chapter_name):
+            shutil.rmtree(chapter_name)
+        # 构建命令
+        command = [
+            sys.executable,
+            getdata_script,
+            "--chapter",
+            chapter_name
+        ]
+        
+        print(f"正在执行命令: {' '.join(command)}")
+        
+        # 调用 getdata.py
+        try:
+            subprocess.run(command, check=True)
+            print(f"{chapter_name} 的 getdata.py 执行成功")
+            
+            # 删除 out/raw 下以 0000_ 或 0001_ 开头的文件
+            raw_dir = os.path.join(current_dir, "out", "raw")
+            if os.path.exists(raw_dir):
+                for filename in os.listdir(raw_dir):
+                    if filename.startswith("0000_paragraphs") or filename.startswith("0001_paragraphs") or filename.startswith("0002_paragraphs") or filename.startswith("0000_pages") or filename.startswith("0001_pages") or filename.startswith("0002_pages"):
+                        file_to_delete = os.path.join(raw_dir, filename)
+                        try:
+                            os.remove(file_to_delete)  # 有些好像又不需要删除
+                            print(f"已删除: {file_to_delete}")
+                        except Exception as e:
+                            print(f"删除 {file_to_delete} 失败: {e}")
+            
+            # 运行 gendatav2.py
+            print(f"正在运行 {chapter_name} 的 gendatav2.py")
+            gendata_command = [sys.executable, gendata_script]
+            subprocess.run(gendata_command, check=True)
+            print(f"{chapter_name} 的 gendatav2.py 执行成功")
+            
+            # 创建目标文件夹
+            target_dir = os.path.join(current_dir, chapter_name)
+            os.makedirs(target_dir, exist_ok=True)
+            print(f"确保目标文件夹存在: {target_dir}")
+            
+            # 移动 out 和 out2
+            for folder_name in ["out", "out2"]:
+                src_dir = os.path.join(current_dir, folder_name)
+                dest_dir = os.path.join(target_dir, folder_name)
+                
+                if os.path.exists(src_dir):
+                    # 如果目标路径已存在同名文件夹，先删除，防止移动时发生嵌套或报错
+                    if os.path.exists(dest_dir):
+                        shutil.rmtree(dest_dir)
+                    shutil.move(src_dir, dest_dir)
+                    print(f"已将 {folder_name} 移动到 {chapter_name} 内")
+                else:
+                    print(f"未找到 {folder_name} 文件夹，跳过")
+                    
+        except subprocess.CalledProcessError as e:
+            print(f"{chapter_name} 的 getdata.py 执行失败，返回码: {e.returncode}")
+        except Exception as e:
+            print(f"发生错误: {e}")
+            
+        # 如果不是最后一个章节，则延迟 3 分钟
+        if i < len(chapters) - 1:
+            print(f"\n{chapter_name} 处理完成。")
+            print("等待 3 分钟 (180秒) 后继续处理下一个章节...")
+            for remaining in range(180, 0, -10):
+                print(f"还剩 {remaining} 秒...")
+                time.sleep(min(10, remaining))
+            print("延迟结束，准备继续。\n")
+
+    print("\n=====================================")
+    print("所有章节处理完成，准备将生成的章节文件夹移动到上一层目录...")
+    print("=====================================")
+    
+    parent_dir = os.path.dirname(current_dir)
+    for chapter_name in chapters:
+        src_dir = os.path.join(current_dir, chapter_name)
+        dest_dir = os.path.join(parent_dir, chapter_name)
+        
+        if os.path.exists(src_dir):
+            try:
+                # 如果目标路径已存在，为了避免 shutil.move 嵌套，可以选择删除或者直接���盖，这里选择覆盖(先删后移)
+                if os.path.exists(dest_dir):
+                    print(f"上一层目录已存在 {chapter_name}，正在删除旧文件夹...")
+                    shutil.rmtree(dest_dir)
+                shutil.move(src_dir, dest_dir)
+                print(f"成功将 {chapter_name} 移动到: {parent_dir}")
+            except Exception as e:
+                print(f"移动 {chapter_name} 时发生错误: {e}")
+        else:
+            print(f"当前目录未找到 {chapter_name} 文件夹，跳过。")
+            
+    print("\n所有任务执行完毕！")
+
+if __name__ == "__main__":
+    main()

ocr/draw_box.py

@@ -0,0 +1,48 @@
+
+import json
+import glob
+from pathlib import Path
+import cv2
+import numpy as np
+
+def draw_box(pth_webp, pth_boxs):
+
+    with open(pth_boxs, encoding='utf-8') as fp:
+        boxs = json.load(fp)
+        fp.close()
+
+    imgData = np.fromfile(pth_webp, dtype=np.uint8)
+    img = cv2.imdecode(imgData, cv2.IMREAD_UNCHANGED)
+
+    if len(img.shape) != 3:  # 转彩图
+        img_color = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)  # DBNet 原版只能处理彩图，这里转一下
+    else:
+        img_color = img.copy()
+
+
+    for box in boxs:
+        charPoly = box["charPoly"]
+        lu = [charPoly["x0"], charPoly["y0"]]
+        ru = [charPoly["x1"], charPoly["y1"]]
+        rd = [charPoly["x2"], charPoly["y2"]]
+        ld = [charPoly["x3"], charPoly["y3"]]
+        points = np.array([lu, ru, rd, ld])
+
+        cv2.polylines(img_color, [points], isClosed=True, color=(  # 多边形，框得比较全
+            100, 0, 255), thickness=2)  # 只画线，不填充
+        
+        # cv2.imshow("box", img_color)
+        # cv2.waitKey(0)
+
+    path_jpg = pth_webp.replace(".webp", ".jpg")
+    cv2.imwrite(path_jpg, img_color)
+
+def do_drawbox():
+    webps = glob.glob('./out2/*.webp', recursive=False)
+    for pth_webp in webps:
+        pth_box = pth_webp.replace(".webp", ".boxs")
+        draw_box(pth_webp, pth_box)
+
+if __name__ == "__main__":
+    do_drawbox()

ocr/gendatav2.py

@@ -0,0 +1,145 @@
+"""
+转换一章数据
+"""
+
+import os
+import shutil
+import json
+import glob
+from pathlib import Path
+import cv2
+import numpy as np
+
+from draw_box import do_drawbox
+
+
+def main():
+    dir_out2 = "out2"
+    if Path(dir_out2).exists():
+        shutil.rmtree(dir_out2)
+
+    if not Path(dir_out2).exists():
+        os.makedirs(dir_out2)
+
+    pageIds_pageNames = {}
+
+    raw_jsons = glob.glob('./out/raw/**/*.json', recursive=True)
+    box_jsons = {}
+    paragraphs = []
+    texts = ""
+    text_pages = {}
+    pages = []
+    for raw_json in raw_jsons:
+        if 'pages_200' in raw_json:
+            with open(raw_json, encoding='utf-8') as fp:
+                pages_json = json.load(fp)
+                page = pages_json["data"]["pages"]
+                pages += page
+                fp.close()
+        if 'paragraphs_200' in raw_json:
+            with open(raw_json, encoding='utf-8') as fp:
+                paragraphs_json = json.load(fp)
+                paragraphs += paragraphs_json["data"]["paragraphs"]
+                fp.close()
+        
+        if 'word_box_200' in raw_json:
+            with open(raw_json, encoding='utf-8') as fp:
+                word_box_json = json.load(fp)
+                box_jsons = {**box_jsons, **word_box_json["data"]["pageId2WordBoxContent"]}
+                fp.close()
+        
+    currPageId = -1
+    lastPageId = -1
+    currPageText = ""
+
+    for i, js in enumerate(paragraphs):
+        startPageId = js["startPageId"]
+        endPageId = js["endPageId"]
+        content = js["content"]
+        content = json.loads(content, strict=False )
+        lines = content["lines"]
+        lastLineNum = -1
+        for line in lines:
+            lineNum = line["lineNum"]
+            lineType = line["lineType"]
+            content_line = line["content"].replace('\ufeff', '').replace("\u3000", " ") #.replace(" ", "")   
+            if 'pagePass' in line:
+                pagePass = line["pagePass"]
+                PageId = pagePass["PageId"]
+
+                if currPageId == -1:
+                    currPageId = PageId
+
+                if currPageText:  # 页切换，保存上一页文本
+                    text_pages[currPageId] = currPageText
+                    currPageText = ""
+                    currPageId = PageId
+
+            
+            currPageText += content_line
+
+            texts += content_line
+
+        if i == len(paragraphs) - 1:
+            if currPageText:  # 最后一个段落，保存上一页文本
+                if currPageId not in text_pages:
+                    text_pages[currPageId] = currPageText
+                    currPageText = ""
+                    currPageId = PageId
+        else:
+            currPageText += " "
+            texts += " "
+        
+
+    # pages = pages_json["data"]["pages"]
+
+    for page in pages:
+        pageId = page["pageId"]
+        pageNum = page["pageNum"]
+        uri = page["uri"]
+        imageName = uri.split("-")[-1]
+        baseName = Path(imageName).stem
+        pth_img = Path("out/images") / imageName
+        if not pth_img.exists():
+            if pth_img.stem in {
+                'TPM0001_00006_00089b', 'TPM0001_00009_00002a', 'TPM0001_00009_00002b', 
+                'TPM0001_00009_00003a', 'TPM0001_00009_00003b', 'TPM0001_00009_00004a', 
+                'TPM0001_00009_00004b', 'TPM0001_00011_00011b', 'TPM0001_00011_00092b'
+            }:
+                continue
+            raise Exception(f"image {imageName} not found")
+        if pageId not in box_jsons:
+            raise Exception(f"pageId {pageId} no box_json")
+
+        pageIds_pageNames[pageId] = baseName
+        boxs = box_jsons[pageId]['wordBoxList']
+        pth_boxs = str( dir_out2 / Path( imageName.replace(".webp", ".boxs") ) )
+
+        shutil.copy(pth_img, str( dir_out2 / Path(imageName) ) )
+        with open(pth_boxs, 'w', encoding='utf-8') as fp:
+            json.dump(boxs, fp, indent=4, ensure_ascii=False)
+            fp.close()
+
+    pth_text_pages = str(Path(dir_out2) / "page_texts.json")
+    with open(pth_text_pages, 'w', encoding='utf-8') as fp:
+        json.dump(text_pages, fp, indent=4, ensure_ascii=False)
+        fp.close()
+
+    pth_pageIds_pageNames = str(Path(dir_out2) / "pageIds_pageNames.json")
+    with open(pth_pageIds_pageNames, 'w', encoding='utf-8') as fp:
+        json.dump(pageIds_pageNames, fp, indent=4, ensure_ascii=False)
+        fp.close()
+
+
+    pth_paragraphs = str(Path(dir_out2) / "paragraphs.json")
+    with open(pth_paragraphs, 'w', encoding='utf-8') as fp:
+        json.dump(paragraphs, fp, indent=4, ensure_ascii=False)
+        fp.close()
+    
+    do_drawbox()
+
+def do_gendata():
+    main()
+
+if __name__ == "__main__":
+    do_gendata()

ocr/getdata.py

@@ -0,0 +1,861 @@
+"""
+取一章数据
+"""
+import argparse
+import base64
+import json
+import re
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional, Tuple
+
+from selenium import webdriver
+from selenium.webdriver.common.by import By
+from selenium.webdriver.support import expected_conditions as EC
+from selenium.webdriver.support.ui import WebDriverWait
+
+TARGET_PATTERNS = {
+    "paragraphs": re.compile(r"/api/ancientlib/read/book/paragraphs/v2(?:\?|$)"),
+    "pages": re.compile(r"/api/ancientlib/read/book/pages/v3/(?:\?|$)"),
+    "word_box": re.compile(r"/api/ancientlib/read/word-box-page-content/m-get/(?:\?|$)"),
+    "loader": re.compile(r"__loader=__session"),
+    "volume": re.compile(r"/api/ancientlib/read/get/volume/(?:\?|$)"),
+}
+
+IMAGE_URL_RE = re.compile(r"https?://[^ ]+(?:\.webp|\.image|\.png|\.jpe?g|\.bmp)(?:\?.*)?$", re.IGNORECASE)
+
+
+def _safe_name(s: str) -> str:
+    s = re.sub(r"[^a-zA-Z0-9._-]+", "_", s)
+    return s[:180] if len(s) > 180 else s
+
+
+def _json_loads_maybe(data: str) -> Optional[Any]:
+    try:
+        return json.loads(data)
+    except Exception:
+        return None
+
+
+def _iter_strings(obj: Any) -> Iterable[str]:
+    if isinstance(obj, str):
+        yield obj
+        return
+    if isinstance(obj, list):
+        for it in obj:
+            yield from _iter_strings(it)
+        return
+    if isinstance(obj, dict):
+        for v in obj.values():
+            yield from _iter_strings(v)
+
+
+def _extract_text_from_json(payload: Any) -> str:
+    if not isinstance(payload, (dict, list)):
+        return ""
+    chunks: List[str] = []
+
+    def walk(o: Any) -> None:
+        if isinstance(o, dict):
+            for k, v in o.items():
+                lk = str(k).lower()
+                if lk in {"text", "content", "paragraph", "para", "value", "word"} and isinstance(v, str):
+                    chunks.append(v)
+                else:
+                    walk(v)
+        elif isinstance(o, list):
+            for it in o:
+                walk(it)
+
+    walk(payload)
+    out = "\n".join(x.strip() for x in chunks if x and x.strip())
+    out = re.sub(r"\n{3,}", "\n\n", out)
+    return out.strip()
+
+
+def _find_image_urls(payload: Any) -> List[str]:
+    urls: List[str] = []
+    for s in _iter_strings(payload):
+        if s.startswith("http") and (".webp" in s or ".image" in s or "/page/" in s):
+            urls.append(s)
+    dedup: List[str] = []
+    seen = set()
+    for u in urls:
+        if u not in seen:
+            seen.add(u)
+            dedup.append(u)
+    return dedup
+
+
+def _maybe_decode_url(s: str) -> List[str]:
+    if not isinstance(s, str) or len(s) < 16:
+        return []
+    out: List[str] = []
+    try:
+        raw = base64.b64decode(s, validate=False)
+        txt = raw.decode("utf-8", errors="ignore")
+    except Exception:
+        return []
+    for m in re.finditer(r"https?://[^\\s\"']+", txt):
+        u = m.group(0)
+        if IMAGE_URL_RE.match(u) and ("byteimg.com" in u or "bytednsdoc.com" in u):
+            out.append(u)
+    return out
+
+
+def _extract_image_urls_from_pages(payload: Any) -> List[str]:
+    urls: List[str] = []
+    if not isinstance(payload, dict):
+        return urls
+    data = payload.get("data")
+    if not isinstance(data, dict):
+        return urls
+    pages = data.get("pages")
+    if not isinstance(pages, list):
+        return urls
+    for p in pages:
+        if not isinstance(p, dict):
+            continue
+        for key in ("picUrl", "thumbUrl"):
+            v = p.get(key)
+            if isinstance(v, str):
+                urls.extend(_maybe_decode_url(v))
+        for v in p.values():
+            if isinstance(v, str) and IMAGE_URL_RE.match(v) and ("byteimg.com" in v or "bytednsdoc.com" in v):
+                urls.append(v)
+    return list(dict.fromkeys(urls))
+
+
+def _guess_page_keys(payload: Any) -> List[Tuple[str, Any]]:
+    hits: List[Tuple[str, Any]] = []
+    if isinstance(payload, dict):
+        for k, v in payload.items():
+            lk = str(k).lower()
+            if lk in {"pageid", "page_id", "page"} and isinstance(v, (str, int)):
+                hits.append((str(k), v))
+            hits.extend(_guess_page_keys(v))
+    elif isinstance(payload, list):
+        for it in payload:
+            hits.extend(_guess_page_keys(it))
+    return hits
+
+
+@dataclass
+class CapturedResponse:
+    kind: str
+    url: str
+    request_id: str
+    status: int
+    mime_type: str
+    body_text: str
+
+    def json(self) -> Optional[Any]:
+        return _json_loads_maybe(self.body_text)
+
+
+def _make_driver(headless: bool) -> webdriver.Chrome:
+    options = webdriver.ChromeOptions()
+    if headless:
+        options.add_argument("--headless=new")
+    options.add_argument("--no-sandbox")
+    options.add_argument("--disable-dev-shm-usage")
+    options.add_argument("--disable-blink-features=AutomationControlled")
+    options.add_argument("user-agent=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36")
+    options.add_experimental_option("excludeSwitches", ["enable-automation"])
+    options.add_experimental_option("useAutomationExtension", False)
+    options.set_capability("goog:loggingPrefs", {"performance": "ALL"})
+    driver = webdriver.Chrome(options=options)
+    driver.execute_cdp_cmd("Network.enable", {})
+    return driver
+
+
+def _drain_network(
+    driver: webdriver.Chrome,
+    tracked: Dict[str, Tuple[str, str, int, str]],
+    ready: List[CapturedResponse],
+    image_urls: List[str],
+    img_dir: Path,
+    max_images: int,
+    img_saved: List[Dict[str, Any]],
+) -> None:
+    try:
+        logs = driver.get_log("performance")
+    except Exception:
+        return
+    for entry in logs:
+        try:
+            msg = json.loads(entry["message"])["message"]
+        except Exception:
+            continue
+        method = msg.get("method")
+        params = msg.get("params", {})
+        if method == "Network.requestWillBeSent":
+            request = params.get("request", {})
+            url = request.get("url", "")
+            if "00015" in url:
+                print(f"[DEBUG] Network.requestWillBeSent for 00015: url={url}")
+            if IMAGE_URL_RE.match(url) and ("byteimg.com" in url or "bytednsdoc.com" in url):
+                image_urls.append(url)
+        elif method == "Network.responseReceived":
+            response = params.get("response", {})
+            url = response.get("url", "")
+            mime_type = str(response.get("mimeType", "") or "")
+            if IMAGE_URL_RE.match(url) and ("byteimg.com" in url or "bytednsdoc.com" in url):
+                image_urls.append(url)
+            request_id = params.get("requestId")
+            if request_id and mime_type.startswith("image/") and IMAGE_URL_RE.match(url) and ("byteimg.com" in url or "bytednsdoc.com" in url):
+                status = int(response.get("status", 0) or 0)
+                tracked[request_id] = ("image", url, status, mime_type)
+                print(f"[DEBUG] Tracked image response: request_id={request_id}, url={url}, status={status}")
+                continue
+            elif request_id and "00015" in url:
+                print(f"[DEBUG] Found 00015 in url but didn't track as image. mime_type={mime_type}, url={url}, request_id={request_id}")
+            kind = None
+            for k, pat in TARGET_PATTERNS.items():
+                if pat.search(url):
+                    kind = k
+                    break
+            if not kind:
+                continue
+            request_id = params.get("requestId")
+            if not request_id:
+                continue
+            status = int(response.get("status", 0) or 0)
+            tracked[request_id] = (kind, url, status, mime_type)
+        elif method == "Network.loadingFinished":
+            request_id = params.get("requestId")
+            if not request_id or request_id not in tracked:
+                continue
+            kind, url, status, mime_type = tracked.pop(request_id)
+            try:
+                body = driver.execute_cdp_cmd("Network.getResponseBody", {"requestId": request_id})
+            except Exception as e:
+                if kind == "image":
+                    print(f"[DEBUG] Failed to get response body for image {url}: {e}")
+                continue
+            if kind == "image":
+                if max_images > 0 and len(img_saved) >= max_images:
+                    print(f"[DEBUG] Max images reached, skipping {url}")
+                    continue
+                raw = body.get("body", "")
+                if not raw:
+                    print(f"[DEBUG] Empty body for image {url}")
+                    continue
+                if body.get("base64Encoded"):
+                    try:
+                        data = base64.b64decode(raw)
+                    except Exception as e:
+                        print(f"[DEBUG] Failed to decode base64 for image {url}: {e}")
+                        continue
+                else:
+                    data = raw.encode("utf-8", errors="ignore")
+                ext = ".bin"
+                if "webp" in (mime_type or "").lower():
+                    ext = ".webp"
+                elif "png" in (mime_type or "").lower():
+                    ext = ".png"
+                elif "jpeg" in (mime_type or "").lower() or "jpg" in (mime_type or "").lower():
+                    ext = ".jpg"
+                
+                page_id = ""
+                m = re.search(r"/page/([^/]+)/", url)
+                if m:
+                    page_id = f"_{m.group(1)}"
+                else:
+                    m2 = re.search(r"1k[a-z0-9]{11}", url)
+                    if m2:
+                        page_id = f"_{m2.group(0)}"
+                
+                # Default file name format
+                file_name = f"{len(img_saved):04d}{page_id}{ext}"
+                
+                # Try to extract actual filename from URL, e.g. SWX0005_00001_00001.webp
+                m_name = re.search(r"-([a-zA-Z0-9_]+\.(?:webp|png|jpe?g|jpg|bmp))", url, re.IGNORECASE)
+                if m_name:
+                    file_name = m_name.group(1)
+                else:
+                    m_name2 = re.search(r"/([^/]+?\.(?:webp|png|jpe?g|jpg|bmp))(?:[?~]|$)", url, re.IGNORECASE)
+                    if m_name2:
+                        name_part = m_name2.group(1)
+                        if "-" in name_part:
+                            file_name = name_part.split("-", 1)[-1]
+                        else:
+                            file_name = name_part
+                
+                out_path = img_dir / file_name
+                try:
+                    out_path.write_bytes(data)
+                    print(f"[DEBUG] Saved image: {file_name} from {url}")
+                except Exception as e:
+                    print(f"[DEBUG] Failed to save image {file_name} from {url}: {e}")
+                    continue
+                img_saved.append({"url": url, "mimeType": mime_type, "path": str(out_path)})
+                continue
+
+            body_text = body.get("body", "")
+            if body.get("base64Encoded"):
+                try:
+                    body_text = base64.b64decode(body_text).decode("utf-8", errors="replace")
+                except Exception:
+                    body_text = ""
+                    
+            if kind == "loader":
+                try:
+                    data = json.loads(body_text)
+                    if "paragraphList" in data:
+                        para_payload = {
+                            "errorCode": 0,
+                            "errorMsg": "",
+                            "data": {
+                                "paragraphs": data["paragraphList"]
+                            }
+                        }
+                        body_text = json.dumps(para_payload, ensure_ascii=False)
+                        kind = "paragraphs"
+                        url = "https://www.shidianguji.com/api/ancientlib/read/book/paragraphs/v2?mock=from_loader"
+                        print(f"[DEBUG] Transformed loader response to paragraphs format")
+                except Exception as e:
+                    print(f"[DEBUG] Failed to parse loader JSON: {e}")
+
+            ready.append(
+                CapturedResponse(
+                    kind=kind,
+                    url=url,
+                    request_id=request_id,
+                    status=status,
+                    mime_type=mime_type,
+                    body_text=body_text,
+                )
+            )
+
+
+def _collect_dom_image_urls(driver: webdriver.Chrome) -> List[str]:
+    try:
+        urls = driver.execute_script(
+            "return Array.from(document.images||[]).map(i=>i.currentSrc||i.src).filter(Boolean);"
+        )
+    except Exception:
+        return []
+    if not isinstance(urls, list):
+        return []
+    out: List[str] = []
+    for u in urls:
+        if isinstance(u, str) and IMAGE_URL_RE.match(u) and ("byteimg.com" in u or "bytednsdoc.com" in u):
+            out.append(u)
+    return list(dict.fromkeys(out))
+
+
+def _try_click_by_text(driver: webdriver.Chrome, text: str, timeout_s: float = 2.5) -> bool:
+    xp = (
+        f"//*[self::button or self::a or @role='button' or self::div]"
+        f"[contains(normalize-space(.), {json.dumps(text, ensure_ascii=False)})]"
+    )
+    try:
+        els = driver.find_elements(By.XPATH, xp)
+    except Exception:
+        return False
+    for el in els[:5]:
+        try:
+            if not el.is_displayed() or not el.is_enabled():
+                continue
+            el.click()
+            return True
+        except Exception:
+            continue
+    return False
+
+
+def _enter_image_mode(driver: webdriver.Chrome) -> None:
+    for t in ("原图", "影印", "图片", "图像", "掃圖", "扫描", "切换"):
+        if _try_click_by_text(driver, t, timeout_s=0):
+            time.sleep(0.8)
+            break
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--book-id", default="HY1540")
+    ap.add_argument("--chapter", default="評論出像水滸傳卷之一_楔子") # 因为目录有两个之四， 不写或者写 [0] 表示点击第一个（默认行为） 写 [1] 表示点击第二个
+    ap.add_argument("--out", default="out")
+    ap.add_argument("--headless", action="store_true")
+    ap.add_argument("--timeout", type=int, default=20000)
+    ap.add_argument("--max-images", type=int, default=1000)
+    args = ap.parse_args()
+
+    out_dir = Path(args.out).resolve()
+    raw_dir = out_dir / "raw"
+    img_dir = out_dir / "images"
+    coord_dir = out_dir / "coords"
+    text_dir = out_dir / "text"
+    for d in (raw_dir, img_dir, coord_dir, text_dir):
+        d.mkdir(parents=True, exist_ok=True)
+
+    driver = _make_driver(headless=args.headless)
+    driver.set_window_size(1400, 900)
+    captured: List[CapturedResponse] = []
+    tracked: Dict[str, Tuple[str, str, int, str]] = {}
+    image_urls: List[str] = []
+    img_saved: List[Dict[str, Any]] = []
+    try:
+        url = f"https://www.shidianguji.com/zh/book/{args.book_id}"
+        driver.get(url)
+        
+        # Drain network to capture the initial chapter data
+        for _ in range(5):
+            time.sleep(1)
+            before = len(captured)
+            _drain_network(driver, tracked, captured, image_urls, img_dir, int(args.max_images), img_saved)
+            if len(captured) == before and any(c.kind in ("paragraphs", "pages") for c in captured):
+                break
+
+        initial_captured = list(captured)
+        initial_img_saved = list(img_saved)
+        initial_image_urls = list(image_urls)
+        
+        captured.clear()
+        img_saved.clear()
+        image_urls.clear()
+        tracked.clear()
+
+        chapter_parts = args.chapter.split("_")
+        current_part_idx = 0
+
+        clicked = False
+        no_scroll_count = 0
+        for scroll_attempts in range(400):
+            raw_target_part = chapter_parts[current_part_idx]
+            
+            # 解析可选的索引，例如 "新刻金瓶梅詞話卷之四[1]" 表示点击第2个匹配的元素（索引从0开始）
+            target_part = raw_target_part
+            target_index = 0
+            m_idx = re.search(r'\[(\d+)\]$', raw_target_part)
+            if m_idx:
+                target_index = int(m_idx.group(1))
+                target_part = raw_target_part[:m_idx.start()]
+
+            part_clicked_in_this_attempt = False
+
+            # 处理常见的异体字，如 "囘" 和 "回"
+            search_parts = [target_part]
+            if '囘' in target_part:
+                search_parts.append(target_part.replace('囘', '回'))
+            if '回' in target_part:
+                search_parts.append(target_part.replace('回', '囘'))
+
+            els = []
+            actual_part = target_part
+            for sp in search_parts:
+                try:
+                    # 寻找目标元素
+                    els = driver.find_elements(By.XPATH, f"//*[contains(normalize-space(.), {json.dumps(sp, ensure_ascii=False)}) and not(.//*[contains(normalize-space(.), {json.dumps(sp, ensure_ascii=False)})])]")
+                    if not els:
+                        els = driver.find_elements(By.LINK_TEXT, sp)
+                    if not els:
+                        els = driver.find_elements(By.PARTIAL_LINK_TEXT, sp)
+                    if els:
+                        actual_part = sp
+                        break
+                except Exception as e:
+                    print(f"[DEBUG] Exception finding '{sp}': {e}")
+                    pass
+
+            target_part = actual_part
+
+            try:
+                # 为了调试，打印找到的元素数量
+                if els:
+                    print(f"[DEBUG] Found {len(els)} elements for '{target_part}'")
+                    for idx_debug, el_debug in enumerate(els):
+                        try:
+                            rect = driver.execute_script("return arguments[0].getBoundingClientRect();", el_debug)
+                            tag_name = el_debug.tag_name
+                            text_content = el_debug.text.strip()
+                            print(f"[DEBUG] Element {idx_debug}: tag={tag_name}, text='{text_content}', rect={rect}")
+                        except Exception as e:
+                            print(f"[DEBUG] Element {idx_debug}: Error getting info: {e}")
+                    
+                seen_y_coords = []
+                for el in els:
+                    try:
+                        # 确保元素可见，但有时候框架中节点在视口外会被认为是 is_displayed() == False，
+                        # 可以先尝试滚动到视图内再判断。
+                        driver.execute_script("arguments[0].scrollIntoView({block: 'center'});", el)
+                        time.sleep(0.3)
+                        
+                        # 某些特殊的折叠列表可能元素高度/宽度为0，我们通过 JS 获取位置和尺寸
+                        rect = driver.execute_script("return arguments[0].getBoundingClientRect();", el)
+                        if rect['width'] <= 0 or rect['height'] <= 0:
+                            continue
+                            
+                        # 根据 Y 坐标去重，同一个位置的多个元素只算作一个匹配项
+                        el_y = rect['top']
+                        is_new_item = True
+                        for seen_y in seen_y_coords:
+                            if abs(seen_y - el_y) < 5:
+                                is_new_item = False
+                                break
+                        
+                        if is_new_item:
+                            seen_y_coords.append(el_y)
+                            
+                        # seen_y_coords 的长度就是当前发现了几个独立的视觉项
+                        # 如果当前项的索引（长度-1）小于目标索引，则跳过
+                        if len(seen_y_coords) - 1 < target_index:
+                            continue
+                            
+                        click_success = False
+                        try:
+                            # 如果元素有特殊的父级可点击区域，有时候点击自身会失败，尝试点击父级
+                            # 在识典古籍中，目录经常是特定的 div 或者 span
+                            el.click()
+                            click_success = True
+                        except Exception:
+                            # Try JS click as fallback if standard click fails
+                            try:
+                                driver.execute_script("arguments[0].click();", el)
+                                click_success = True
+                            except Exception:
+                                pass
+                                
+                        # 如果点击成功了，我们需要检查它是不是实际上展开了。
+                        # 因为在有些 UI 中，点击一个已经被选中的层级会把它折叠起来。
+                        if click_success:
+                            print(f"[DEBUG] Clicked '{target_part}' successfully.")
+                            # 等待展开动画
+                            time.sleep(1.5)
+                            
+                            # 判断是否成功展开：如果是父节点，我们需要能看到它的子节点
+                            if current_part_idx < len(chapter_parts) - 1:
+                                raw_next_part = chapter_parts[current_part_idx + 1]
+                                next_part = raw_next_part
+                                m_next_idx = re.search(r'\[(\d+)\]$', raw_next_part)
+                                if m_next_idx:
+                                    next_part = raw_next_part[:m_next_idx.start()]
+                                
+                                next_search_parts = [next_part]
+                                if '囘' in next_part:
+                                    next_search_parts.append(next_part.replace('囘', '回'))
+                                if '回' in next_part:
+                                    next_search_parts.append(next_part.replace('回', '囘'))
+                                    
+                                next_els = []
+                                actual_next_part = next_part
+                                for nsp in next_search_parts:
+                                    next_els = driver.find_elements(By.XPATH, f"//*[contains(normalize-space(.), {json.dumps(nsp, ensure_ascii=False)}) and not(.//*[contains(normalize-space(.), {json.dumps(nsp, ensure_ascii=False)})])]")
+                                    if not next_els:
+                                        next_els = driver.find_elements(By.LINK_TEXT, nsp)
+                                    if not next_els:
+                                        next_els = driver.find_elements(By.PARTIAL_LINK_TEXT, nsp)
+                                    if next_els:
+                                        actual_next_part = nsp
+                                        break
+                                        
+                                is_next_visible = False
+                                for next_el in next_els:
+                                    try:
+                                        if driver.execute_script("var rect = arguments[0].getBoundingClientRect(); return rect.width > 0 && rect.height > 0;", next_el):
+                                            is_next_visible = True
+                                            break
+                                    except:
+                                        pass
+                                        
+                                if not is_next_visible:
+                                    print(f"[DEBUG] '{target_part}' was clicked but '{actual_next_part}' is still not visible. It might have been collapsed. Clicking again.")
+                                    # 再点一次将其展开
+                                    try:
+                                        el.click()
+                                    except:
+                                        driver.execute_script("arguments[0].click();", el)
+                                    time.sleep(1.5)
+                                    
+                            part_clicked_in_this_attempt = True
+                            if current_part_idx == len(chapter_parts) - 1:
+                                clicked = True
+                            else:
+                                current_part_idx += 1
+                                no_scroll_count = 0
+                            break
+                    except Exception as e:
+                        print(f"[DEBUG] Exception while interacting with '{target_part}': {e}")
+                        continue
+            except Exception as e:
+                print(f"[DEBUG] Exception finding '{target_part}': {e}")
+                pass
+                
+            if clicked:
+                break
+                
+            if part_clicked_in_this_attempt:
+                continue
+                
+            # 如果没找到或没点击成功，滚动所有可滚动的 div，尝试让章节列表显示出来
+            try:
+                scrolled = driver.execute_script('''
+                    var els = document.querySelectorAll("div, ul, main, nav, section, aside");
+                    var scrolledAny = false;
+                    for (var i = 0; i < els.length; i++) {
+                        var d = els[i];
+                        if (d.scrollHeight > d.clientHeight && window.getComputedStyle(d).overflowY !== "hidden") {
+                            var before = d.scrollTop;
+                            d.scrollBy(0, 250);
+                            if (d.scrollTop > before) {
+                                scrolledAny = true;
+                            }
+                        }
+                    }
+                    return scrolledAny;
+                ''')
+                if not scrolled:
+                    no_scroll_count += 1
+                    if no_scroll_count >= 20:
+                        print(f"[DEBUG] Reached the bottom of the list (tried 20 times). {target_part} not found.")
+                        break
+                else:
+                    no_scroll_count = 0
+            except Exception:
+                pass
+            time.sleep(0.5)
+        
+        if clicked:
+            print(f"[DEBUG] Clicked {args.chapter}, waiting for new network data...")
+            # Wait until we see new 'paragraphs' or 'pages' in captured
+            for _ in range(15):
+                time.sleep(1)
+                _drain_network(driver, tracked, captured, image_urls, img_dir, int(args.max_images), img_saved)
+                if any(c.kind in ("paragraphs", "pages") for c in captured):
+                    print(f"[DEBUG] New data captured for {args.chapter}")
+                    break
+        else:
+            print(f"[DEBUG] {args.chapter} not clicked (not found).")
+            
+        has_new_data = any(c.kind in ("paragraphs", "pages") for c in captured)
+        
+        if not has_new_data:
+            if clicked:
+                print(f"[DEBUG] No new data loaded after clicking {args.chapter}. Using initial chapter data.")
+            else:
+                print(f"[DEBUG] Using initial chapter data because {args.chapter} was not found.")
+            captured.extend(initial_captured)
+            img_saved.extend(initial_img_saved)
+            image_urls.extend(initial_image_urls)
+        else:
+            print(f"[DEBUG] Successfully loaded new chapter {args.chapter}. Merging missing data types from initial chapter data.")
+            for kind in ("pages", "word_box", "volume", "paragraphs"):
+                if not any(c.kind == kind for c in captured):
+                    print(f"[DEBUG] Inheriting {kind} from initial capture.")
+                    captured.extend([c for c in initial_captured if c.kind == kind])
+            
+            # Keep initial images as they might be part of the same volume
+            img_saved.extend(initial_img_saved)
+            image_urls.extend(initial_image_urls)
+
+        time.sleep(0.8)
+        _enter_image_mode(driver)
+
+        deadline = time.time() + max(10, int(args.timeout))
+        last_activity = time.time()
+        no_activity_count = 0
+        print(f"Starting loop, deadline in {deadline - time.time()} seconds")
+        
+        # Try to find a scrollable container
+        from selenium.webdriver.common.keys import Keys
+        
+        while time.time() < deadline:
+            before_cap = len(captured)
+            before_img = len(img_saved)
+            _drain_network(driver, tracked, captured, image_urls, img_dir, int(args.max_images), img_saved)
+            if len(captured) != before_cap or len(img_saved) != before_img:
+                print(f"Activity! captured: {len(captured)} (+{len(captured)-before_cap}), img_saved: {len(img_saved)} (+{len(img_saved)-before_img})")
+                last_activity = time.time()
+                no_activity_count = 0
+            
+            # Send PAGE_DOWN to body
+            try:
+                driver.find_element(By.TAG_NAME, "body").send_keys(Keys.PAGE_DOWN)
+            except Exception:
+                pass
+                
+            # Click next page if possible
+            try:
+                # The user indicated: <div class="vik-toolbox-btn click act" type="page-next">
+                next_btn = driver.find_element(By.XPATH, "//div[@type='page-next' or contains(@class, 'page-next') or contains(text(), '下一张')]")
+                # Add a small delay to prevent clicking too fast which might skip pages
+                # Check if we're not loading something
+                if next_btn.is_displayed():
+                    # Check if button is disabled (has disable class)
+                    cls = next_btn.get_attribute("class") or ""
+                    if "disable" not in cls:
+                        # Find current page text to ensure we only click once page changes
+                        page_text_el = driver.find_elements(By.XPATH, "//div[contains(text(), '/')]")
+                        curr_text = ""
+                        for p in page_text_el:
+                            if "/" in p.text:
+                                curr_text = p.text
+                                break
+                        driver.execute_script("arguments[0].click();", next_btn)
+                        # Wait for page to change
+                        for _ in range(10):
+                            time.sleep(0.2)
+                            new_text = ""
+                            for p in driver.find_elements(By.XPATH, "//div[contains(text(), '/')]"):
+                                if "/" in p.text:
+                                    new_text = p.text
+                                    break
+                            if new_text != curr_text:
+                                break
+            except Exception:
+                pass
+                
+            # Also try to scroll window and any potential scrollable divs
+            driver.execute_script('''
+                window.scrollBy(0, 900);
+                var divs = document.querySelectorAll("div");
+                for (var i = 0; i < divs.length; i++) {
+                    if (divs[i].scrollHeight > divs[i].clientHeight && window.getComputedStyle(divs[i]).overflowY !== "hidden") {
+                        divs[i].scrollBy(0, 900);
+                    }
+                }
+            ''')
+            
+            time.sleep(0.6)
+            image_urls.extend(_collect_dom_image_urls(driver))
+            if time.time() - last_activity > 15:
+                no_activity_count += 1
+                print(f"No activity for 15 seconds (count: {no_activity_count}).")
+                if no_activity_count >= 3:
+                    print("Too many consecutive periods of no activity, breaking loop.")
+                    break
+                    
+                # Check if next button is actually disabled
+                try:
+                    next_btn = driver.find_element(By.XPATH, "//div[@type='page-next' or contains(@class, 'page-next') or contains(text(), '下一张')]")
+                    cls = next_btn.get_attribute("class") or ""
+                    if "disable" in cls:
+                        print("Reached end of chapter (next button disabled).")
+                        break
+                    else:
+                        print("Next button not disabled, but no network activity. Try clicking again.")
+                        driver.execute_script("arguments[0].click();", next_btn)
+                        last_activity = time.time() - 10 # give it 5 more seconds before checking again
+                        continue
+                except Exception:
+                    print("Could not find next button, breaking loop.")
+                    break
+                break
+        print(f"Loop finished. time.time() < deadline: {time.time() < deadline}")
+        
+        for i, c in enumerate(captured):
+            ext = "json" if "json" in (c.mime_type or "") or c.body_text.strip().startswith("{") else "txt"
+            name = _safe_name(f"{i:04d}_{c.kind}_{c.status}_{c.url}")
+            (raw_dir / f"{name}.{ext}").write_text(c.body_text, encoding="utf-8", errors="replace")
+
+        img_urls = list(dict.fromkeys([u for u in image_urls if u.startswith("http")]))
+        pages_payloads = [c.json() for c in captured if c.kind == "pages" and c.status == 200]
+        
+        page_id_to_hash = {}
+        for p in pages_payloads:
+            if p is not None:
+                img_urls.extend(_extract_image_urls_from_pages(p))
+                # Build pageId -> hash mapping
+                if isinstance(p, dict) and isinstance(p.get("data"), dict):
+                    pages_list = p["data"].get("pages", [])
+                    if isinstance(pages_list, list):
+                        for page in pages_list:
+                            if isinstance(page, dict):
+                                pid = str(page.get("pageId", ""))
+                                uri = str(page.get("uri", ""))
+                                if pid and uri:
+                                    m = re.search(r"/page/([^/]+)/", uri)
+                                    if m:
+                                        page_id_to_hash[pid] = m.group(1)
+                                    else:
+                                        m2 = re.search(r"1k[a-z0-9]{11}", uri)
+                                        if m2:
+                                            page_id_to_hash[pid] = m2.group(0)
+
+        img_urls = list(dict.fromkeys(img_urls))
+
+        coords: Dict[str, Any] = {}
+        for c in captured:
+            if c.kind != "word_box" or c.status != 200:
+                continue
+            payload = c.json()
+            if payload is None:
+                continue
+            
+            page_id = None
+            # Extract pageId directly from pageId2WordBoxContent if available
+            if isinstance(payload, dict) and isinstance(payload.get("data"), dict):
+                content = payload["data"].get("pageId2WordBoxContent")
+                if isinstance(content, dict) and content:
+                    page_id = str(next(iter(content.keys())))
+            
+            if not page_id:
+                page_keys = _guess_page_keys(payload)
+                for _, v in page_keys:
+                    if isinstance(v, (str, int)):
+                        page_id = str(v)
+                        break
+                        
+            hash_suffix = ""
+            if page_id and page_id in page_id_to_hash:
+                hash_suffix = f"_{page_id_to_hash[page_id]}"
+                
+            key = f"wordbox_{len(coords):04d}{hash_suffix}"
+            coords[key] = payload
+            (coord_dir / f"{_safe_name(key)}.json").write_text(
+                json.dumps(payload, ensure_ascii=False),
+                encoding="utf-8",
+                errors="replace",
+            )
+
+        paragraph_texts: List[str] = []
+        for c in captured:
+            if c.kind != "paragraphs" or c.status != 200:
+                continue
+            payload = c.json()
+            if payload is None:
+                continue
+            t = _extract_text_from_json(payload)
+            if t:
+                paragraph_texts.append(t)
+        full_text = "\n\n".join(paragraph_texts).strip()
+        if not full_text and coords:
+            fallback_chunks: List[str] = []
+            for v in coords.values():
+                t = _extract_text_from_json(v)
+                if t:
+                    fallback_chunks.append(t)
+            full_text = "\n\n".join(fallback_chunks).strip()
+
+        if full_text:
+            (text_dir / "text.txt").write_text(full_text, encoding="utf-8", errors="replace")
+
+        summary = {
+            "bookId": args.book_id,
+            "chapter": args.chapter,
+            "captured": [
+                {"kind": c.kind, "status": c.status, "mimeType": c.mime_type, "url": c.url}
+                for c in captured
+            ],
+            "imageUrls": img_urls[: max(0, int(args.max_images))],
+            "imagesSaved": img_saved,
+            "coords_keys": list(coords.keys()),
+            "out": str(out_dir),
+        }
+        (out_dir / "summary.json").write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")
+        return 0
+    finally:
+        try:
+            driver.quit()
+            from gendatav2 import do_gendata
+            do_gendata()
+        except Exception:
+            pass
+
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
+

ocr/readme.txt

@@ -0,0 +1,247 @@
+
+see huggingface_echodict\typst_jpm
+
+https://www.shidianguji.com/zh/book/HY1540
+
+
+uv venv --python 3.10 --seed --clear
+
+.venv/Scripts/pip.exe install -r requirements.txt --trusted-host pypi.org --trusted-host files.pythonhosted.org --trusted-host pypi.python.org
+
+.venv/Scripts/python.exe
+  
+
+python getdata.py
+  先修改第几章，运行下载文本和坐标画框转换格式
+
+duckdb + ducklake + readest + typst
+
+针对竖排/繁体/异体字与旧书噪声，选型
+
+**推荐选型（你的场景：竖排 + 繁体/异体字 + 旧书噪声）**
+
+- 首选检测：`CRAFT` 或 `DBNet++`
+- 首选识别：`SVTR-LCNet` / `PARSeq`（中文词表可扩）
+- 实战最稳组合：`CRAFT(检测) + SVTR(识别)` 或 `DBNet++(检测) + PARSeq(识别)`
+
+**为什么这样选**
+
+- `CRAFT` 优势：
+  - 对字符级与复杂排版（竖排、密集、非规则间距）通常更稳。
+  - 对旧书里断笔、脏污、背景纹理干扰有较好容错。
+  - 缺点是后处理和部署复杂度较高，速度一般。
+
+- `DBNet++` 优势：
+  - 工业成熟、训练和部署相对标准化。
+  - 对低对比、模糊页表现好。
+  - 缺点是天然偏“文本区域”，若你要严格“单字框”，需更细后处理或字符级训练策略。
+
+- `YOLO` 在你这个特定场景的定位：
+  - 可作为强基线（快、易训），但在旧书噪声+极小字+竖排密集时，通常上限不如 `CRAFT/DBNet++` 体系稳定。
+
+**针对“繁体/异体字”最关键点（比模型更重要）**
+
+- 字典必须扩展到繁体与常见异体字（否则识别上限被词表卡死）。
+- 训练数据要覆盖旧字形、破损字、弱印刷样式。
+- 识别模型建议用可扩词表的 Transformer 系（`PARSeq`、`SVTR`）而不是仅固定简体词集模型。
+
+**给你一个直接可执行的决策**
+
+- 要“最高鲁棒性”（不太在乎部署复杂）：`CRAFT + PARSeq/SVTR`
+- 要“精度和工程平衡”：`DBNet++ + SVTR`
+- 要“快速上线先跑通”：`YOLOv8字符检测 + SVTR`，后续再切 `CRAFT/DBNet++`
+
+**一句话结论**
+- 你的场景优先从 `DBNet++` 和 `CRAFT` 二选一做检测，识别端用 `SVTR/PARSeq` 并扩繁体异体词表；这比单纯换更大检测器更能提升最终效果。
+
+
+https://github.com/wkentaro/labelme
+  文本标注
+
+https://borninfreedom.github.io/posts/2024/07/blog-post-1/ 
+    论文解析——Character Region Awareness for Text Detection，字符级文本检测CRAFT算法
+
+https://arxiv.org/abs/1912.04561  A Feasible Framework for Arbitrary-Shaped Scene Text Recognition
+    https://github.com/zhang0jhon/AttentionOCR 代码
+    ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text Competition的冠军
+    它具有以下几个优点：中英文通用、准确性高、代码开源和有预训练模型。简而言之，亲测好用！
+        https://zhuanlan.zhihu.com/p/138589087
+
+
+https://www.shuge.org/meet/topic/78721/
+    https://www.ancientbooks.cn/
+    书格
+
+https://wenyuan.aliyun.com/
+    汉典重光
+
+
+https://www.shidianguji.com/zh/book/SWX0005 这网页的顶部有一个按钮 “原本影像”，点一下能出图片。最右边是正文，最左边是章节目录。正文里每一个字符都有一组坐标值能与对应的图片中的字符框对应。看看它在前端是怎样建立对应关系的，相应 api 是什么，怎么下载相关数据，包括文字，坐标，原图
+    https://solo.trae.ai/
+
+<div class="vik-toolbox-btn click act" type="page-next">  这里有一个"下一张" 按钮，点击会切换到下一张图片，看看它是怎来得到下一张图片的，把后面的图片都下载下来
+
+
+A）最快：Network 里 “Copy as fetch” → 粘贴到 Console 执行
+对每个接口都能用（paragraphs / word-box）：
+
+打开书页（比如 https://www.shidianguji.com/zh/book/SWX0005 进入阅读）
+F12 → Network
+过滤 paragraphs/v2 或 word-box-page-content/m-get
+点开一条真实请求（确保 Preview/Response 有数据，不是 0 bytes）
+右键该请求 → Copy → Copy as fetch
+粘贴到 Console 回车执行
+返回的就是 JSON（then(r=>r.json())...）
+如果你遇到 bdturing-verify（滑块验证），先在网页上完成验证后再复制那条“成功返回数据”的请求。
+
+(() => {
+  window.__cap = { fetch: [], xhr: [] };
+
+  // hook fetch
+  const _fetch = window.fetch;
+  window.fetch = async function(input, init = {}) {
+    try {
+      const url = typeof input === "string" ? input : input.url;
+      const method = (init.method || "GET").toUpperCase();
+      const body = init.body;
+      if (url.includes("/api/ancientlib/read/")) {
+        window.__cap.fetch.push({ t: Date.now(), url, method, init: { ...init, body } });
+        console.log("[CAP fetch]", method, url, body ? "(has body)" : "");
+      }
+    } catch (e) {}
+    return _fetch.apply(this, arguments);
+  };
+
+  // hook XHR
+  const _open = XMLHttpRequest.prototype.open;
+  const _send = XMLHttpRequest.prototype.send;
+  XMLHttpRequest.prototype.open = function(method, url) {
+    this.__cap_info = { t: Date.now(), method, url, body: null };
+    return _open.apply(this, arguments);
+  };
+  XMLHttpRequest.prototype.send = function(body) {
+    try {
+      if (this.__cap_info && String(this.__cap_info.url).includes("/api/ancientlib/read/")) {
+        this.__cap_info.body = body;
+        window.__cap.xhr.push(this.__cap_info);
+        console.log("[CAP xhr]", this.__cap_info.method, this.__cap_info.url, body ? "(has body)" : "");
+      }
+    } catch (e) {}
+    return _send.apply(this, arguments);
+  };
+
+  console.log("抓包器已安装：现在请在页面里翻页/滚动/点击“原本影像”等触发接口请求。");
+})();
+
+
+function lastReq(match) {
+  const all = [...window.__cap.fetch, ...window.__cap.xhr]
+    .filter(x => x.url && x.url.includes(match))
+    .sort((a,b) => b.t - a.t);
+  return all[0];
+}
+
+async function replay(req) {
+  if (!req) throw new Error("没抓到请求，请先触发一次页面请求。");
+  const url = req.url;
+  const method = (req.method || "GET").toUpperCase();
+  const body = req.init?.body ?? req.body ?? null;
+
+  const res = await fetch(url, {
+    method,
+    // 不要乱加 Origin/Referer（浏览器会自己带）；关键是带登录态：
+    credentials: "include",
+    headers: { "content-type": "application/json" },
+    body
+  });
+
+  const text = await res.text();
+  if (!text) {
+    console.warn("响应为空，可能触发滑块验证(bdturing)。请先在网页上完成验证后再抓一次。");
+    return null;
+  }
+  try { return JSON.parse(text); } catch (e) {
+    console.log("非 JSON 文本前200：", text.slice(0, 200));
+    throw e;
+  }
+}
+
+// 1） 拿 paragraphs
+(async () => {
+  const req = lastReq("/book/paragraphs/v2");
+  console.log("使用请求：", req?.method, req?.url);
+  const data = await replay(req);
+  console.log("paragraphs:", data);
+  window.__paragraphs = data; // 缓存
+})();
+
+// 2） 拿 word-box
+(async () => {
+  const req = lastReq("/word-box-page-content/m-get/");
+  console.log("使用请求：", req?.method, req?.url);
+  const data = await replay(req);
+  console.log("wordbox:", data);
+  window.__wordbox = data; // 缓存
+})();
+
+
+
+https://ocr.kandianguji.com/ocr_api
+
+token：8be3d75a-8a66-4b5d-9047-d4895e4c8000
+
+email：13788325535
+
+请求参数：
+
+token：您所申请的API Token，点此申请 必传
+
+email：申请API Token的账号，看典古籍网站的注册账号（参数名为email，手机号注册的传手机号） 必传
+
+image：需要识别的古籍图像，base64编码后的字符串类型，您可以在此处转换您的图像为base64编码 必传
+
+char_ocr：是否进行单字符检测识别，不检测文本行，只检测图像上的字符，文本行顺序可能会出错；布尔类型；默认值：False
+
+det_mode：文字内容排版样式，目前有三种可选：auto（自动识别）、sp（竖向排版）、hp（横向排版）；字符串类型，默认值：auto
+
+image_size：识别前图像尺寸调整，图像越小识别速度越快，0为不调整，设置指定值将按照设置对图像最长边进行等比例调整；整数类型，默认值：0
+
+return_position：是否返回文本行坐标信息和字符坐标信息；布尔类型，默认值：False
+
+return_choices：是否返回每个字符的其它候选字；布尔类型，默认值：False
+
+version：指定识别算法版本；字符串类型，可选：default（v1标准版本）、beta（古籍语序优化版本）、v2（最新版本），默认值：default
+
+det_layout（v2）：是否开启版面识别（对图像上的内容进行判断是否是正文、页眉页脚/版心等），对分栏式、多栏式文档识别效果较好；布尔类型（开启True、关闭False），默认值：False
+
+only_plain_text（v2）：是否只识别返回正文内容，仅当版面识别开启时生效，不识别页眉/页脚/版心等；布尔类型（开启True、关闭False），默认值：False
+
+return_layout（v2）：是否返回版面信息，仅当版面识别开启时生效；布尔类型（开启True、关闭False），默认值：False
+
+auto_insert_space（v2）：按照字符间距自动插入空格；布尔类型（开启True、关闭False），默认值：False
+
+hp_line_words_angel（v2）：指定横排句子文字排序方向；字符串类型（从左到右left2right、从右到左right2left），默认值：left2right
+
+sp_line_words_angel（v2）：指定竖排句子文字排序方向；字符串类型（从上到下top2bottom、从下到上bottom2top），默认值：top2bottom
+
+请求方式：
+
+POST请求，请求体可以为Form Data或JSON两种方式均可接受
+
+
+
+
+duckdb + ducklake + readest + typst see huggingface_echodict\typst_hlm\ocr\readme.txt
+
+https://aistudio.baidu.com/projectdetail/4438534?channelType=0&channel=0 真实 ocr 项目经验
+
+
+https://huggingface.co/datasets/ByteDance/AncientDoc 字节的数据集
+
+![img](深入理���神经网络：从逻辑回归到CNN.assets/b5ec48e89a6402434386c3479c641e7b.png)
+
+https://aistudio.baidu.com/datasetdetail/165369 真可以下载的 **古籍数据集**
+
+https://huggingface.co/datasets/Teklia/CASIA-HWDB2-line
+
+https://github.com/esun-ai/traditional-chinese-text-recogn-dataset

ocr/requirements.txt

@@ -0,0 +1,4 @@
+requests
+selenium==4.40.0
+numpy<2.0.0
+opencv-python==4.10.0.84